Single-stream single-stage target tracking method and system based on dual softmax attention

By adopting dual softmax attention mechanism and ViT model in the target tracker, combining feature diversity preservation and proxy bias, the problem of high computational complexity of Transformer tracker is solved, and efficient target tracking performance and simplified framework structure are achieved. , improves the accuracy and robustness of the tracker.

CN118691852BActive Publication Date: 2024-11-05ZHONGBIN WENTOU (TIANJIN) TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411192503.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-08-28
Publication Date
2024-11-05
Estimated Expiration
2044-08-28

AI Technical Summary

Technical Problem

The Transformer-based target tracker calculates the similarity between each pixel by self-attention, resulting in high computational complexity, affecting real-time, making it difficult to achieve efficient tracking in real scenes.

Method used

Using a single-stream single-stage target tracking method with dual softmax attention, replacing the encoder in the ViT model, using the dual softmax attention mechanism for feature learning, combining the ViT model for unified feature extraction and fusion, reducing the computational complexity and improving the model Expressive ability and introduce feature diversity preservation operations and proxy bias to enhance location information utilization.

Benefits of technology

It realizes the maintenance of good tracking performance while reducing computing costs, simplifies the dual-stream and dual-stage tracking framework, and improves the accuracy and robustness of the tracker.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118691852B_ABST
    Figure CN118691852B_ABST
Patent Text Reader

Abstract

The present invention proposes a single-stream single-stage target tracking method and system based on dual softmax attention. The method is based on the ViT model, and the encoder in the ViT model is replaced with an encoder based on dual softmax attention to obtain a target tracking model, and the template and the image of the search area are input into the target tracking model. Through the encoder of the dual softmax attention, unified feature learning is performed on the template image and the search area image to establish a free information flow between the template and the search area, and feature information with target orientation is obtained; and the result is sent to the prediction head for classification and regression operations respectively to obtain the target tracking result. The present invention combines the characteristics of the linear attention mechanism and the softmax attention mechanism, and has low computational complexity and high model expression ability, so that the tracker can achieve good tracking performance while reducing the computational cost.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of computer vision and image processing, and in particular to a single-stream single-stage target tracking method and system based on dual softmax attention. Background Art

[0002] Object tracking is one of the main research directions in the field of computer vision. After the initial state of the target in a given video sequence, the position of the target is calibrated using a bounding box in subsequent video frames to achieve continuous positioning and state estimation of the target. Currently, object tracking is widely used in public security, autonomous driving, intelligent robots, human-computer intelligent interaction, medical diagnosis and other fields.

[0003] In real-world target tracking, the target is subject to various challenging factors, mainly because the target is a non-specific object and the target tracker cannot pre-train or model the target in advance. Moreover, during the tracking process, the target will undergo various changes and be disturbed by the surrounding environment, such as target disappearance, appearance changes, background interference, rapid movement, etc. These factors pose great challenges to the tracking algorithm.

[0004] In recent years, Transformer-based trackers have achieved encouraging tracking performance. Transformer can be used to model the relationship between various basic visual elements, including pixel-pixel, object-pixel, and object-object. Convolution is a local operation, and a convolution layer usually only models the relationship between adjacent pixels. Transformer is a global operation, and the Transformer layer can model the relationship between all pixels, and the two can complement each other well. Convolution can be regarded as a template matching, using the same template to filter different locations in the image. The attention unit in Transformer is an adaptive filter, and this adaptive computing module has strong modeling capabilities. Compared with CNN, Transformer is more competitive in some visual tasks and shows better performance.

[0005] However, Transformer-based trackers usually bring huge computational burdens due to the self-attention calculation of the similarity between each pixel, which affects the real-time performance of the tracker. Therefore, how to reduce the computational complexity of attention becomes a key factor in improving the speed of Transformer-based trackers. Summary of the invention

[0006] In view of the above situation, the main purpose of the present invention is to propose a single-stream single-stage target tracking method and system based on dual softmax attention to solve the above technical problems.

[0007] The present invention proposes a single-stream single-stage target tracking method based on dual softmax attention, the method comprising the following steps:

[0008] Step 1: Based on the ViT model, the encoder in the ViT model is replaced with an encoder based on dual softmax attention to obtain a target tracking model, and the template and the image of the search area are input into the target tracking model;

[0009] Step 2: Preprocess the images of the template and the search area and input them into the encoder based on dual softmax attention;

[0010] Step 3: The encoder based on dual softmax attention uses two softmax attentions in each attention head to perform attention calculation on the input features, and then performs unified feature learning on the template image and the search area image to establish a free information flow between the template and the search area, realize mutual guidance between the template and the search area, and obtain goal-oriented feature information;

[0011] Step 4: Repeat step 3 several times in an iterative manner to dynamically model the relationship between the template and the search area features to obtain target-oriented features;

[0012] Step 5: Send the target-oriented features to the prediction head for classification and regression operations, and obtain the center position, bounding box size and position deviation of the target to be tracked. The center position is adjusted according to the deviation, and the final tracking box is determined according to the size of the bounding box to obtain the target tracking result.

[0013] Step 6: Repeat steps 2 to 4 using a large-scale data set to pre-train the target tracking model to obtain a pre-trained target tracking model, and use the pre-trained target tracking model to perform target tracking to obtain accurate target tracking results.

[0014] The present invention also proposes a single-stream single-stage target tracking system based on dual softmax attention, wherein the system applies the single-stream single-stage target tracking method based on dual softmax attention as described above, and the system comprises:

[0015] Building blocks for:

[0016] Based on the ViT model, the encoder in the ViT model is replaced with an encoder based on dual softmax attention to obtain a target tracking model, and the template and the image of the search area are input into the target tracking model;

[0017] Learning modules for:

[0018] The encoder based on dual softmax attention uses two softmax attentions in each attention head to calculate the input features, and then performs unified feature learning on the template image and the search area image to establish a free information flow between the template and the search area, realize mutual guidance between the template and the search area, and obtain goal-oriented feature information;

[0019] Extraction module for:

[0020] The images of the template and the search area are preprocessed and input into the encoder based on double softmax attention;

[0021] The learning module is iteratively repeated several times to dynamically model the relationship between the template and the search area features to obtain target-oriented features.

[0022] Compute module for:

[0023] The target-oriented features are sent to the prediction head for classification and regression operations to obtain the center position, bounding box size and position deviation of the target to be tracked. The center position is adjusted according to the deviation, and the final tracking box is determined according to the size of the bounding box to obtain the target tracking result.

[0024] Pre-trained modules for:

[0025] Pre-training the target tracking model using a large-scale data set to obtain a pre-trained target tracking model;

[0026] Tracking module for:

[0027] Use the pre-trained target tracking model to track the target to obtain accurate target tracking results.

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] 1. The dual softmax attention mechanism successfully integrates the characteristics of the linear attention mechanism and the softmax attention mechanism, while having lower computational complexity and higher model expression ability, enabling the tracker to achieve good tracking performance while reducing computational costs.

[0030] 2. As a generalized linear attention mechanism, the dual softmax attention mechanism has the problem of insufficient feature diversity. Therefore, the present invention introduces a feature diversity preservation operation to solve this problem. In addition, in order to better utilize the position information, a carefully designed proxy bias is provided for the dual softmax attention mechanism.

[0031] 3. The present invention adopts a single-stream single-stage tracking framework, which perfectly couples feature extraction and feature fusion together, simplifies the previous two-stream two-stage tracking framework, and achieves good tracking performance.

[0032] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description or learned through embodiments of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 Flow chart of the single-stream single-stage target tracking method based on dual softmax attention proposed by the present invention;

[0034] Figure 2 The overall architecture of the single-stream single-stage target tracking method based on dual softmax attention proposed in this invention;

[0035] Figure 3 Schematic diagram of the dual softmax attention module of the present invention;

[0036] Figure 4 This is a schematic diagram of the structure of the single-stream single-stage target tracking system based on dual softmax attention proposed in the present invention. DETAILED DESCRIPTION

[0037] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0038] These and other aspects of the embodiments of the present invention will be apparent with reference to the following description and accompanying drawings. In these descriptions and accompanying drawings, some specific implementations of the embodiments of the present invention are specifically disclosed to represent some ways of implementing the principles of the embodiments of the present invention, but it should be understood that the scope of the embodiments of the present invention is not limited thereto.

[0039] See also Figure 1 and Figure 2 , this embodiment provides a single-stream single-stage target tracking method based on dual softmax attention, the method comprising the following steps:

[0040] Step 1: Based on the ViT model, the encoder in the ViT model is replaced with an encoder based on dual softmax attention to obtain a target tracking model, and the template and the image of the search area are input into the target tracking model;

[0041] Step 2: Preprocess the images of the template and the search area and input them into the encoder based on dual softmax attention;

[0042] In step 2, the method for preprocessing the template and the image of the search area specifically includes the following steps:

[0043] Divide the input template and search area image pair into small blocks;

[0044] Position encoding of small patches of the image of the template and the search area;

[0045] Flatten the position-encoded small blocks to form serialized image blocks;

[0046] The serialized image patches of the template and the search area are concatenated as the input of the double softmax attention based encoder.

[0047] Step 3: The encoder based on dual softmax attention uses two softmax attentions in each attention head to perform attention calculation on the input features, and then performs unified feature learning on the template image and the search area image to establish a free information flow between the template and the search area, realize mutual guidance between the template and the search area, and obtain feature information with goal-orientedness, thereby extracting target-oriented features and avoiding the loss of discriminative information, and obtaining template and search area features with strong discriminative power;

[0048] In step 3, the method of using two softmax attentions in each attention head to calculate the attention of the input features based on the dual softmax attention encoder specifically includes the following steps:

[0049] First, the original input features are normalized by one layer to stabilize the input distribution of each layer and speed up the convergence of the model. The corresponding process has the following relationship:

[0050] ;

[0051] ;

[0052] in, represents the original input features, Represents the original input features The mean of Represents the original input features The standard deviation of and Represent the learnable parameters for scaling and translating the normalized output, Representation layer normalization operation, Represents normalized output.

[0053] The normalized output is fed into the dual softmax attention module to perform relationship modeling, obtain features with long-term dependencies, and obtain attention output;

[0054] The attention output is layer-normalized and then fed into a multi-layer perceptron to capture the complex relationship between the original input features and enhance the model's expressiveness. The output of the multi-layer perceptron is then residually connected to the attention output to alleviate the gradient vanishing problem and obtain the output features of the encoder based on dual softmax attention. The corresponding process has the following relationship:

[0055] ;

[0056] in, represents a multi-layer perceptron, represents the attention output, Indicates the encoder output.

[0057] See also Figure 3 Preferably, the normalized output is fed into a dual softmax attention module to perform relationship modeling and obtain features with long-term dependencies. The method for obtaining the attention output specifically includes the following steps:

[0058] Linear projection is performed on the normalized output to obtain the query vector, key vector, and value vector required to calculate the attention. The corresponding process has the following relationship:

[0059] ;

[0060] in, Represent the query vector, key vector and value vector respectively, Respectively used to obtain The linear projection matrix.

[0061] Adaptive average pooling is performed on the query vector to obtain the proxy token, and then double softmax attention calculation is performed to obtain the intermediate features. The corresponding process has the following relationship:

[0062] ;

[0063] ;

[0064] in, Represents a proxy token, represents the intermediate feature, represents multi-head dual softmax attention calculation, represents adaptive average pooling;

[0065] In order to retain more original low-level feature information, residual connections are used to add the intermediate features obtained by attention calculation to the original input features to obtain the attention output. The corresponding process has the following relationship:

[0066] .

[0067] Preferably, the multi-head dual softmax attention calculation process specifically includes the following steps:

[0068] The proxy token As a query, with the key vector Sum value vector Perform softmax attention calculation to aggregate proxy features from all values ;

[0069] The proxy features are used as the value matrix in the second softmax attention, and the proxy tokens are used as the key matrix. The second softmax attention calculation is performed between the query vector, the proxy tokens, and the proxy features to obtain the final output, the output of a single attention head, and then the outputs of multiple attention heads are concatenated to obtain the intermediate features. The corresponding process has the following relationship:

[0070] ;

[0071] ;

[0072] ;

[0073] in, represents the transpose operation, Respectively represent The query vector, proxy token, key vector, and value vector in the attention heads. represents the softmax activation function, Indicates The output of an attention head is Represents double softmax attention calculation.

[0074] It can be seen from the above generalized form of combined linear attention that the dual softmax attention can be regarded as a generalized linear attention, where the calculation formula of the generalized linear attention is expressed as:

[0075] ;

[0076] in, Represents a mapping function.

[0077] It can be seen from the calculation formula of generalized linear attention that when the mapping function is softmax, the overall form of the two is the same, and the dual softmax attention of the present invention can be regarded as a generalized linear attention. Therefore, the present invention uses two softmax operations to convert the original self-attention into a linear form, and can have the advantages of both at the same time, can reduce its computational complexity to linear, and at the same time can have the efficient expression ability of self-attention.

[0078] Preferably, a feature diversity preservation operation is also included, and the feature diversity preservation operation includes a depthwise separable convolution and a residual connection operation. For each attention head, the value vector is sequentially subjected to a depthwise separable convolution operation and a residual connection operation, and then combined with the double softmax attention calculation result of the corresponding attention head to restore feature diversity. The corresponding process has the following relationship:

[0079] ;

[0080] in, Represents a feature diversity preserving operation.

[0081] Preferably, a proxy bias is also included, and the proxy bias is combined with the double softmax attention calculation to better utilize the position information of the target. The method of combining the proxy bias with the double softmax attention calculation specifically includes the following steps:

[0082] Given three learnable bias components, the three bias components are column bias, row bias and block bias;

[0083] Initialize and perform bilinear interpolation operations on the three bias components respectively;

[0084] Add the three bias components after bilinear interpolation to obtain the proxy bias;

[0085] The softmax attention operation of each attention head in the dual softmax attention calculation process is added to the proxy bias for combination. The corresponding process has the following relationship:

[0086] ;

[0087] in, They represent the corresponding proxy biases in the two softmax attention calculations.

[0088] Step 4: Repeat step 3 several times in an iterative manner to dynamically model the relationship between the template and the search area features to obtain target-oriented features;

[0089] Step 5: Send the target-oriented features to the prediction head for classification and regression operations, and obtain the center position, bounding box size and position deviation of the target to be tracked. The center position is adjusted according to the deviation, and the final tracking box is determined according to the size of the bounding box to obtain the target tracking result.

[0090] Step 6: Repeat steps 2 to 4 using a large-scale data set to pre-train the target tracking model to obtain a pre-trained target tracking model, and use the pre-trained target tracking model to perform target tracking to obtain accurate target tracking results.

[0091] The present invention proposes a single-stream single-stage target tracking method based on dual softmax attention, which includes: constructing a network model integrating feature extraction and feature fusion based on dual softmax attention and ViT model. Through the ViT model based on dual softmax attention, unified feature learning is performed on the template image and the search area image to establish a free information flow between the template and the search area, thereby extracting target-oriented features and avoiding the loss of discriminant information, and obtaining template and search area features with strong discriminative power; pre-training the network model based on dual softmax attention and ViT model; using the pre-trained network model, feature learning is performed on the template image and the search area image, and the feature learning results are input into the center positioning prediction head to predict the center position of the target and the size of the bounding box to obtain the final tracking result. The present invention integrates feature extraction and feature fusion together through dual softmax attention and ViT model, which effectively improves the accuracy and robustness of the proposed tracker.

[0092] Please refer to Figure 4 This embodiment also provides a single-stream single-stage target tracking system based on dual softmax attention, wherein the system applies the single-stream single-stage target tracking method based on dual softmax attention as described above, and the system includes:

[0093] Building blocks for:

[0094] Based on the ViT model, the encoder in the ViT model is replaced with an encoder based on dual softmax attention to obtain a target tracking model, and the template and the image of the search area are input into the target tracking model;

[0095] Learning modules for:

[0096] The encoder based on dual softmax attention uses two softmax attentions in each attention head to calculate the input features, and then performs unified feature learning on the template image and the search area image to establish a free information flow between the template and the search area, realize mutual guidance between the template and the search area, and obtain goal-oriented feature information;

[0097] Extraction module for:

[0098] The images of the template and the search area are preprocessed and input into the encoder based on double softmax attention;

[0099] The learning module is iteratively repeated several times to dynamically model the relationship between the template and the search area features to obtain target-oriented features.

[0100] Compute module for:

[0101] The target-oriented features are sent to the prediction head for classification and regression operations to obtain the center position, bounding box size and position deviation of the target to be tracked. The center position is adjusted according to the deviation, and the final tracking box is determined according to the size of the bounding box to obtain the target tracking result.

[0102] Pre-trained modules for:

[0103] Pre-training the target tracking model using a large-scale data set to obtain a pre-trained target tracking model;

[0104] Tracking module for:

[0105] Use the pre-trained target tracking model to perform target tracking to obtain accurate target tracking results.

[0106] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0107] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0108] The above-mentioned embodiments only express several implementation methods of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the patent of the present invention. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

Claims

1. A single-stream single-stage target tracking method based on dual softmax attention, It is characterized in that The method comprises the following steps: Step 1: Based on the ViT model, the encoder in the ViT model is replaced with an encoder based on dual softmax attention to obtain a target tracking model, and the template and the image of the search area are input into the target tracking model; Step 2: Preprocess the images of the template and the search area and input them into the encoder based on dual softmax attention; Step 3: The encoder based on dual softmax attention uses two softmax attentions in each attention head to perform attention calculation on the input features, and then performs unified feature learning on the template image and the search area image to establish a free information flow between the template and the search area, realize mutual guidance between the template and the search area, and obtain goal-oriented feature information; Step 4: Repeat step 3 several times in an iterative manner to dynamically model the relationship between the template and the search area features to obtain target-oriented features; Step 5: Send the target-oriented features to the prediction head for classification and regression operations, and obtain the center position, bounding box size and position deviation of the target to be tracked. The center position is adjusted according to the deviation, and the final tracking box is determined according to the size of the bounding box to obtain the target tracking result. Step 6: Repeat steps 2 to 4 using a large-scale data set to pre-train the target tracking model to obtain a pre-trained target tracking model, and use the pre-trained target tracking model to perform target tracking to obtain accurate target tracking results; In step 3, the method of using two softmax attentions in each attention head to calculate the attention of the input features based on the dual softmax attention encoder specifically includes the following steps: First, the original input features are normalized by one layer to stabilize the input distribution of each layer and speed up the convergence of the model. The corresponding process has the following relationship: ; ; in, represents the original input features, Represents the original input features The mean of Represents the original input features The standard deviation of and Represent the learnable parameters for scaling and translating the normalized output, Representation layer normalization operation, represents the normalized output; The normalized output is fed into a dual softmax attention module to perform relationship modeling, obtain features with long-term dependencies, and obtain attention output; The attention output is layer-normalized and then fed into a multi-layer perceptron to capture the complex relationship between the original input features and enhance the model's expressiveness. The output of the multi-layer perceptron is then residually connected to the attention output to alleviate the gradient vanishing problem and obtain the output features of the encoder with double softmax attention. The corresponding process has the following relationship: ; in, represents a multi-layer perceptron, represents the attention output, Indicates encoder output; The normalized output is sent to the dual softmax attention module to perform relationship modeling and obtain features with long-term dependencies. The method for obtaining the attention output specifically includes the following steps: Linear projection is performed on the normalized output to obtain the query vector, key vector, and value vector required to calculate the attention. The corresponding process has the following relationship: ; ; ; in, Represent the query vector, key vector and value vector respectively, Respectively used to obtain The linear projection matrix of Adaptive average pooling is performed on the query vector to obtain the proxy token, and then double softmax attention calculation is performed to obtain the intermediate features. The corresponding process has the following relationship: ; ; in, Represents a proxy token, represents the intermediate feature, represents multi-head dual softmax attention calculation, represents adaptive average pooling; In order to retain more original low-level feature information, residual connections are used to add the intermediate features obtained by attention calculation to the original input features to obtain the attention output. The corresponding process has the following relationship: ; The multi-head dual softmax attention calculation process specifically includes the following steps: Take the proxy token as the query, perform softmax attention calculation with the key vector and value vector, and aggregate the proxy features from all values; The proxy features are used as the value matrix in the second softmax attention, and the proxy tokens are used as the key matrix. The second softmax attention calculation is performed between the query vector, the proxy tokens, and the proxy features to obtain the final output, the output of a single attention head, and then the outputs of multiple attention heads are concatenated to obtain the intermediate features. The corresponding process has the following relationship: ; ; ; in, represents the transpose operation, denote the query vector, proxy token, key vector, and value vector in the i-th attention head, respectively. represents the softmax activation function, represents the output of the i-th attention head, Represents double softmax attention calculation.

2. The single-stream single-stage target tracking method based on dual softmax attention according to claim 1, It is characterized in that In step 2, the method for preprocessing the template and the image of the search area specifically includes the following steps: Divide the input template and search area image pair into small blocks; Position encoding of small patches of the image of the template and the search area; Flatten the position-encoded small blocks to form serialized image blocks; The serialized image patches of the template and the search area are concatenated as the input of the double softmax attention based encoder.

3. The single-stream single-stage target tracking method based on dual softmax attention according to claim 1, It is characterized in that It also includes feature diversity preservation operations, which include depthwise separable convolution and residual connection operations. For each attention head, the value vector is sequentially subjected to depthwise separable convolution and residual connection operations, and then combined with the double softmax attention calculation results of the corresponding attention head to restore feature diversity. The corresponding process has the following relationship: ; in, Represents a feature diversity preserving operation.

4. The single-stream single-stage target tracking method based on dual softmax attention according to claim 3, It is characterized in that It also includes a proxy bias, which is combined with the double softmax attention calculation to better utilize the position information of the target. The method of combining the proxy bias with the double softmax attention calculation specifically includes the following steps: Given three learnable bias components, the three bias components are column bias, row bias and block bias; Initialize and perform bilinear interpolation operations on the three bias components respectively; Add the three bias components after bilinear interpolation to obtain the proxy bias; The softmax attention operation of each attention head in the dual softmax attention calculation process is added to the proxy bias for combination. The corresponding process has the following relationship: ; in, They represent the corresponding proxy biases in the two softmax attention calculations.

5. A single-stream single-stage target tracking system based on dual softmax attention, It is characterized in that The system applies the single-stream single-stage target tracking method based on dual softmax attention according to any one of claims 1 to 4, and the system comprises: Building blocks for: Based on the ViT model, the encoder in the ViT model is replaced with an encoder based on dual softmax attention to obtain a target tracking model, and the template and the image of the search area are input into the target tracking model; Learning modules for: The encoder based on dual softmax attention uses two softmax attentions in each attention head to calculate the input features, and then performs unified feature learning on the template image and the search area image to establish a free information flow between the template and the search area, realize mutual guidance between the template and the search area, and obtain goal-oriented feature information; Extraction module for: The images of the template and the search area are preprocessed and input into the encoder based on double softmax attention; The learning module is iteratively repeated several times to dynamically model the relationship between the template and the search area features to obtain target-oriented features. Compute module for: The target-oriented features are sent to the prediction head for classification and regression operations to obtain the center position, bounding box size and position deviation of the target to be tracked. The center position is adjusted according to the deviation, and the final tracking box is determined according to the size of the bounding box to obtain the target tracking result. Pre-trained modules for: Pre-training the target tracking model using a large-scale data set to obtain a pre-trained target tracking model; Tracking module for: Use the pre-trained target tracking model to track the target to obtain accurate target tracking results.